As already identified, the ST works fine for smallish datasets, but consumes prodigious amounts of memory for the millions of small anonymous arrays, when used on large datasets.

In addition to that, you are compounding things by creating multiple, huge, stack-based lists in the following line:

    for $key( map{ $_ -> [0] } sort{ $a->[1] cmp $b->[1] || $a->[2] cm
+p $b->[2] } map{ [ $_,(split)[0],(split)[2] ] } keys %hash )
##          ^list1            ^list2                                  
+          ^list3                               ^list4
##                                                                    
+                       ^^^^^^ dups ^^^^^
[download]

And duplicating effort by performing the same split twice for every record.

If you moved your code to using a GRT (see A brief tutorial on Perl's native sorting facilities.) it would probably use about 1/4 the memory, and run in about 1/10th the time.